letta-ai/letta-code

What it solves

Letta Code provides a framework for building stateful AI agents that maintain long-term memory, identity, and a sense of experience over time. Unlike standard LLM interactions that often reset or lose context, these agents are designed to evolve by programmatically updating their own memory, skills, and prompts, allowing them to work proactively on tasks over long horizons.

How it works

The system acts as a harness for LLMs, enabling them to manage their own state. Agents can "dream" (reflect on their experiences), learn new skills, and rewrite their own system prompts. It uses a file-system-like approach (MemFS) to track context via Git and supports multi-agent orchestration, where agents can call other agents as subagents. Users can interact with these agents via CLI, web interfaces, or messaging platforms like Slack and Discord, with the option to sync agent state across multiple remote computers using Letta Cloud.

Who it’s for

It is for developers and researchers building autonomous, long-running AI agents that require persistent memory and the ability to adapt their behavior based on past interactions and evolving task requirements.

Highlights

• Self-Improvement: Agents can autonomously update their own memory blocks and learn new skills.
• Multi-Machine Support: Sync and run agents across different computers, including cloud VMs and local machines.
• Extensible Integrations: Connects to Slack, Telegram, Discord, and custom channels, with support for custom automation hooks.
• Research-Backed: Built by the creators of MemGPT, focusing on continual learning and memory-augmented AI.